Chatbots that handle Telugu and Hindi, not just English

The problem is switching, not translation

Multilingual support is usually sold as a language toggle: pick English or Telugu, get that version of the bot. Real enquiries do not work that way. People switch language mid-conversation without thinking about it, often mid-sentence, and they mix scripts — Telugu words typed in Roman letters are the norm on a phone keyboard.

A bot built around a toggle handles the first message and loses the second. Multilingual support matters most on parent and patient enquiries, which is exactly where switching is most common, because those are the conversations people care enough about to say precisely.

Which languages are worth building for

For our clients that usually means Telugu and Hindi alongside English, and not much else. Adding languages your customers do not use costs money in testing and in every future change to the knowledge base.

The honest way to decide is to read your last two hundred WhatsApp enquiries and count. Most businesses find the answer is two languages, sometimes three, and that the third is rarer than they assumed.

Preserving meaning, not flattening it

The failure mode of a multilingual AI layer is that everything comes out as slightly formal English regardless of what went in. Register and script both get lost, and the reply reads as though nobody was listening.

We have built against this before. The Safalya & Praneeth wedding platform lets guests write a blessing in any of ten Indian languages, including Telugu and Hindi, and refines the words while keeping the meaning, the cultural register and the script intact rather than translating them into generic English. The same constraint applies to an enquiry bot: the answer should come back in the language and register the person used.

What to specify before anyone builds it

A chatbot scope described in features cannot be checked at the end. One described in conversations can.

• The specific questions it must handle, in each language, written out
• What happens when it does not know: silence, a guess and an escalation are three different products
• Where an escalated conversation lands, and who is accountable for answering it
• Which channels it serves — website, WhatsApp, social — and whether one bot serves all of them
• How you change an answer after launch, without raising a request

What it takes to run

Build runs about three weeks end to end: discovery, training on your content, channel and CRM integration, then launch. Deployment covers your website, WhatsApp and social channels, with appointment booking against your real calendar and availability rules, and conversation analytics so you can see what people actually asked.

Running cost is driven by model usage at your volume, which is a question worth answering before signing rather than after the first invoice.

When not to build one

If your enquiry volume is low enough that one person answers everything within the hour, a bot adds a layer without removing work. If the questions genuinely require judgement rather than recall, a bot will guess, and a wrong answer in a parent's own language is worse than no answer at all.

We will say so on a first call rather than take the project.